[PATCH net-next 1/2] tcp: add net.ipv4.tcp_l3mdev_exempt_ports sysctl

From: Kalash Nainwal via B4 Relay

Date: Fri Nov 20 2026 - 17:47:40 EST


From: Kalash Nainwal <kalash@xxxxxxxxxx>

With tcp_l3mdev_accept=1, an unbound listener also matches connections
that arrive in a VRF. If a VRF-bound listener exists on the same port
and both use SO_REUSEPORT, the lookup returns whichever is first in the
listen bucket, not the VRF-bound one. Since AF_INET and AF_INET6
reuseport sockets are inserted at opposite ends of the bucket, the same
setup can pick the VRF-bound listener for IPv4 and the unbound one for
IPv6.

This breaks, for example, a BGP daemon with a TCP-MD5 key only on its
VRF-bound listener: IPv6 peers in the VRF reach the unbound listener
and the session comes up without MD5.

Add net.ipv4.tcp_l3mdev_exempt_ports, a per-netns port list in the
same format as ip_local_reserved_ports. For a listed port, an unbound
socket does not match connections that arrive in a VRF, so only a
VRF-bound listener can take them. Other ports and default-VRF traffic
are not affected. The default is empty.

Signed-off-by: Kalash Nainwal <kalash@xxxxxxxxxx>
---
Documentation/networking/ip-sysctl.rst | 19 +++++++++++++++++++
include/net/ip.h | 14 ++++++++++++++
include/net/netns/ipv4.h | 1 +
net/ipv4/inet_hashtables.c | 5 +++++
net/ipv4/sysctl_net_ipv4.c | 14 ++++++++++++++
net/ipv6/inet6_hashtables.c | 4 ++++
6 files changed, 57 insertions(+)

diff --git a/Documentation/networking/ip-sysctl.rst b/Documentation/networking/ip-sysctl.rst
index 3cc80d9b024..1de0ee5c0d5 100644
--- a/Documentation/networking/ip-sysctl.rst
+++ b/Documentation/networking/ip-sysctl.rst
@@ -619,6 +619,25 @@ tcp_l3mdev_accept - BOOLEAN

Default: 0 (disabled)

+tcp_l3mdev_exempt_ports - list of comma separated ranges
+ TCP ports for which tcp_l3mdev_accept is ignored. When
+ tcp_l3mdev_accept is enabled, a listen socket that is not bound to
+ an L3 master device also matches connections that arrive in an L3
+ domain (e.g., a VRF). For a port in this list, such a socket does
+ not match these connections; only a listen socket bound to the L3
+ master device does. Connections that do not arrive in an L3 domain
+ are not affected. This applies to both IPv4 and IPv6.
+
+ This is useful when the same port has both an unbound listener and
+ a listener bound to a VRF, and VRF connections must reach the
+ VRF-bound one (for example, when only it has TCP-MD5 or TCP-AO
+ keys configured).
+
+ The format is the same as ip_local_reserved_ports, e.g.
+ "179,5000-5010". Writing to the file replaces the current list.
+
+ Default: empty
+
tcp_low_latency - BOOLEAN
This is a legacy option, it has no effect anymore.

diff --git a/include/net/ip.h b/include/net/ip.h
index 5035b313a26..c6581080a9a 100644
--- a/include/net/ip.h
+++ b/include/net/ip.h
@@ -396,6 +396,20 @@ static inline bool inet_port_requires_bind_service(struct net *net, unsigned sho
}
#endif

+/* Ports listed in net.ipv4.tcp_l3mdev_exempt_ports behave as if
+ * tcp_l3mdev_accept were 0: an unbound listener does not match traffic
+ * received on an L3 master device slave. Applies to IPv4 and IPv6.
+ */
+static inline bool tcp_port_l3mdev_exempt(const struct net *net,
+ unsigned short port)
+{
+#ifdef CONFIG_SYSCTL
+ if (net->ipv4.sysctl_tcp_l3mdev_exempt_ports)
+ return test_bit(port, net->ipv4.sysctl_tcp_l3mdev_exempt_ports);
+#endif
+ return false;
+}
+
__be32 inet_current_timestamp(void);

/* From inetpeer.c */
diff --git a/include/net/netns/ipv4.h b/include/net/netns/ipv4.h
index 3efbc7f4b4f..899d90f34a7 100644
--- a/include/net/netns/ipv4.h
+++ b/include/net/netns/ipv4.h
@@ -271,6 +271,7 @@ struct netns_ipv4 {

#ifdef CONFIG_SYSCTL
unsigned long *sysctl_local_reserved_ports;
+ unsigned long *sysctl_tcp_l3mdev_exempt_ports;
int sysctl_ip_prot_sock;
#endif

diff --git a/net/ipv4/inet_hashtables.c b/net/ipv4/inet_hashtables.c
index 860145d810e..063f6dd00a6 100644
--- a/net/ipv4/inet_hashtables.c
+++ b/net/ipv4/inet_hashtables.c
@@ -362,6 +362,11 @@ static inline int compute_score(struct sock *sk, const struct net *net,

if (!inet_sk_bound_dev_eq(net, sk->sk_bound_dev_if, dif, sdif))
return -1;
+
+ if (!sk->sk_bound_dev_if && sdif &&
+ tcp_port_l3mdev_exempt(net, hnum))
+ return -1;
+
score = sk->sk_bound_dev_if ? 2 : 1;

if (sk->sk_family == PF_INET)
diff --git a/net/ipv4/sysctl_net_ipv4.c b/net/ipv4/sysctl_net_ipv4.c
index 9df93ad219d..83d0030512a 100644
--- a/net/ipv4/sysctl_net_ipv4.c
+++ b/net/ipv4/sysctl_net_ipv4.c
@@ -840,6 +840,13 @@ static const struct ctl_table ipv4_net_table[] = {
.mode = 0644,
.proc_handler = proc_do_large_bitmap,
},
+ {
+ .procname = "tcp_l3mdev_exempt_ports",
+ .data = &init_net.ipv4.sysctl_tcp_l3mdev_exempt_ports,
+ .maxlen = 65536,
+ .mode = 0644,
+ .proc_handler = proc_do_large_bitmap,
+ },
{
.procname = "ip_no_pmtu_disc",
.data = &init_net.ipv4.sysctl_ip_no_pmtu_disc,
@@ -1708,10 +1715,16 @@ static __net_init int ipv4_sysctl_init_net(struct net *net)
if (!net->ipv4.sysctl_local_reserved_ports)
goto err_ports;

+ net->ipv4.sysctl_tcp_l3mdev_exempt_ports = kzalloc(65536 / 8, GFP_KERNEL);
+ if (!net->ipv4.sysctl_tcp_l3mdev_exempt_ports)
+ goto err_l3mdev_ports;
+
proc_fib_multipath_hash_set_seed(net, 0);

return 0;

+err_l3mdev_ports:
+ kfree(net->ipv4.sysctl_local_reserved_ports);
err_ports:
unregister_net_sysctl_table(net->ipv4.ipv4_hdr);
err_reg:
@@ -1729,6 +1742,7 @@ static __net_exit void ipv4_sysctl_exit_net(struct net *net)
unregister_net_sysctl_table(net->ipv4.ipv4_hdr);
kfree(table);
kfree(net->ipv4.sysctl_local_reserved_ports);
+ kfree(net->ipv4.sysctl_tcp_l3mdev_exempt_ports);
}

static __net_initdata struct pernet_operations ipv4_sysctl_ops = {
diff --git a/net/ipv6/inet6_hashtables.c b/net/ipv6/inet6_hashtables.c
index 1eea584d80d..43c52b56e8b 100644
--- a/net/ipv6/inet6_hashtables.c
+++ b/net/ipv6/inet6_hashtables.c
@@ -141,6 +141,10 @@ static inline int compute_score(struct sock *sk, const struct net *net,
if (!inet_sk_bound_dev_eq(net, sk->sk_bound_dev_if, dif, sdif))
return -1;

+ if (!sk->sk_bound_dev_if && sdif &&
+ tcp_port_l3mdev_exempt(net, hnum))
+ return -1;
+
score = sk->sk_bound_dev_if ? 2 : 1;
if (READ_ONCE(sk->sk_incoming_cpu) == raw_smp_processor_id())
score++;

--
2.43.0